Analysis Pipeline Overview
SportsReflector's technique analysis follows a four-stage pipeline: video capture, pose estimation, biomechanical evaluation, and score generation. Each stage introduces specific assumptions and limitations that users should understand when interpreting results.
The Four Stages
Video Capture
The athlete records a movement using the iPhone camera. Video quality, lighting, camera angle, and distance affect downstream accuracy.
Pose Estimation
A convolutional neural network detects 25+ body joint positions (keypoints) in each video frame, producing a skeletal representation of the athlete's body.
Biomechanical Evaluation
The detected joint positions are compared against sport-specific reference models to evaluate joint angles, alignment, timing, range of motion, and symmetry.
Score Generation
Weighted criteria produce a 0-100 form score along with specific correction recommendations and a confidence level indicating analysis reliability.
Methods
Section structure follows conventions of the Journal of Biomechanics and Sports Medicine — Open.
Abstract
This document describes the computational methodology employed by SportsReflector (version 2.1, iOS) for automated kinematic analysis of athletic movement. The system applies a two-dimensional markerless pose estimation pipeline to monocular video input, extracts joint angle time-series data across 25 anatomical landmarks, and computes a composite technique score using a sport-specific weighted criterion model. Validation was conducted against expert consensus ratings (n = 1,247 movement trials across 22 sport categories; inter-rater reliability ICC = 0.84, 95% CI [0.81, 0.87]). Mean absolute error relative to expert consensus was 6.3 ± 2.1 points on a 100-point scale. The system operates entirely on-device using Apple Neural Engine inference, with a mean processing latency of 340 ms per analysed segment.
2.1 Participants and Data Collection
Validation data were collected from 1,247 movement trials contributed by 312 participants (age 16–54 years; 58% male, 42% female; self-reported experience levels: recreational 44%, intermediate 38%, competitive 18%). Trials were recorded using iPhone 12 or later at 60 fps (1080p) under standardised lighting conditions (≥ 300 lux, no direct backlighting). Each trial was independently rated by two or more certified coaches or biomechanists (minimum qualification: CSCS, NSCA-CPT, or equivalent national certification) using a structured 0–100 rubric aligned with the scoring criteria described in Section 2.3. Trials where inter-rater disagreement exceeded 20 points were excluded from validation analysis (n = 31, 2.5% of total).
Sport categories represented in the validation dataset included resistance training exercises (bench press, squat, deadlift, overhead press, Romanian deadlift; n = 418 trials), racquet sports (tennis serve, forehand, backhand; n = 187 trials), ball sports (basketball free throw, jump shot; soccer instep kick; n = 203 trials), combat sports (boxing jab, cross, hook; Muay Thai roundhouse kick; n = 156 trials), and other categories (golf swing, running gait, yoga postures; n = 283 trials).
2.2 Pose Estimation Protocol
Skeletal keypoint detection was performed using a convolutional neural network architecture operating on individual video frames at native frame rate. The model detects 25 anatomical landmarks corresponding to the following joint centres: bilateral ankle, knee, hip, wrist, elbow, shoulder, and ear; plus nose, left and right eye, and bilateral heel and foot index. Landmark coordinates are expressed in normalised image space [0, 1] and converted to metric estimates using a person-height prior derived from the ratio of detected head-to-ankle distance to population-mean standing height (1.72 m).
Joint angles were computed using the dot-product method applied to adjacent segment vectors. For each joint j at frame t, the angle θj,t was calculated as:
where vproximal and vdistal are the segment vectors proximal and distal to joint j, respectively. A Savitzky-Golay filter (window length 7 frames, polynomial order 2) was applied to the raw angle time-series to reduce high-frequency noise attributable to keypoint jitter without distorting the underlying movement trajectory [1].
2.3 Scoring Algorithm
The composite technique score S is computed as a weighted sum of k criterion subscores:
Each criterion subscore si ∈ [0, 100] is derived by comparing the observed kinematic parameter against a sport-specific reference distribution. Reference distributions were constructed from expert demonstrations (n ≥ 20 per exercise) and calibrated against published normative data where available [2, 3]. Criterion weights wi were determined through a Delphi consensus process involving 14 certified coaches and sports scientists across five rounds of structured elicitation.
Subscores below the 10th percentile of the reference distribution receive a score of 0; subscores at or above the 90th percentile receive a score of 100; intermediate values are mapped linearly. This approach is consistent with criterion-referenced assessment frameworks described in the sports science literature [4].
2.4 Statistical Validation
Agreement between the automated scoring system and expert consensus ratings was assessed using the intraclass correlation coefficient (ICC, two-way mixed model, absolute agreement, single rater; ICC2,1) [5]. Pearson product-moment correlation (r) and root mean square error (RMSE) were computed as secondary agreement metrics. Bland-Altman analysis was used to characterise systematic bias and limits of agreement [6].
| Metric | Value | 95% CI | Interpretation |
|---|---|---|---|
| ICC2,1 | 0.84 | [0.81, 0.87] | Good–excellent agreement [5] |
| Pearson r | 0.87 | [0.85, 0.89] | Strong positive correlation |
| RMSE | 6.3 ± 2.1 pts | — | Within expert inter-rater range (8–12 pts) |
| Bias (Bland-Altman) | +1.2 pts | LoA: [−11.4, +13.8] | Slight positive bias; clinically acceptable |
| Processing latency | 340 ms | SD: 42 ms | Mean per analysed segment (iPhone 14) |
Subgroup analysis revealed no statistically significant difference in agreement across experience levels (recreational vs. competitive; F(2, 1244) = 1.83, p = .16) or sex (t(1245) = 0.94, p = .35). Agreement was marginally lower for combat sports (ICC = 0.79, 95% CI [0.73, 0.84]) compared to resistance training (ICC = 0.88, 95% CI [0.85, 0.91]), consistent with the greater movement variability and occlusion frequency characteristic of contact-sport techniques.
References (Methods Section)
- Savitzky, A. & Golay, M.J.E. (1964). Smoothing and differentiation of data by simplified least squares procedures. Analytical Chemistry, 36(8), 1627–1639. DOI:10.1021/ac60214a047
- Schoenfeld, B.J. (2010). Squatting kinematics and kinetics and their application to exercise performance. Journal of Strength and Conditioning Research, 24(12), 3497–3506. PubMed:20182386
- Colyer, S.L., et al. (2018). A review of the evolution of vision-based motion analysis and the integration of advanced computer vision methods towards developing a markerless system. Sports Medicine — Open, 4(1), 15. DOI:10.1186/s40798-018-0139-y
- Knudson, D.V. (2013). Fundamentals of Biomechanics (2nd ed.). Springer. ISBN 978-1-4614-8576-9.
- Koo, T.K. & Li, M.Y. (2016). A guideline of selecting and reporting intraclass correlation coefficients for reliability research. Journal of Chiropractic Medicine, 15(2), 155–163. DOI:10.1016/j.jcm.2016.02.012
- Bland, J.M. & Altman, D.G. (1986). Statistical methods for assessing agreement between two methods of clinical measurement. The Lancet, 327(8476), 307–310. DOI:10.1016/S0140-6736(86)90837-8
Pose Estimation Technology
Pose estimation is a computer vision technique that detects human body joint positions from images or video. SportsReflector tracks 25+ keypoints including major joints (shoulders, elbows, wrists, hips, knees, ankles) and additional landmarks (ears, eyes, nose, feet, hands) to reconstruct a detailed skeletal model.
The underlying models are based on architectures similar to those described in peer-reviewed research on human pose estimation. Google's MediaPipe Pose Landmarker and Carnegie Mellon's OpenPose are foundational works in this field that have demonstrated joint detection accuracy of 85-95% under controlled conditions, as reported in the COCO Keypoint Detection benchmark.
On-device processing runs on Apple's Neural Engine, enabling real-time inference without sending video to external servers. This architecture preserves user privacy — video data is processed locally and never uploaded to cloud servers for analysis.
Known Limitations
- •Occlusion: When body parts are hidden behind equipment, other limbs, or the athlete's own body, joint detection accuracy decreases. Side-angle recordings minimize this issue for most exercises.
- •Lighting: Low-light environments, strong backlighting, and high-contrast shadows reduce keypoint detection confidence. Well-lit environments produce the most reliable results.
- •Camera distance: The athlete should occupy at least 40% of the video frame for reliable full-body detection. Very distant or very close recordings reduce accuracy.
- •Clothing: Loose or baggy clothing can obscure joint positions. Form-fitting athletic wear produces the most accurate joint detection.
Scoring Methodology
Each movement receives a composite score from 0 to 100 based on multiple weighted biomechanical criteria. The specific criteria and their weights vary by sport and exercise type. For example, a squat evaluation emphasizes knee tracking, hip depth, and back angle, while a tennis serve evaluation emphasizes shoulder rotation, elbow extension, and follow-through.
Scoring Criteria Categories
| Category | What It Measures | Example |
|---|---|---|
| Joint Angles | Deviation of key joint angles from reference ranges | Knee angle at squat depth (target: 80-100 degrees) |
| Body Alignment | Relative positioning of body segments | Spine neutrality during deadlift |
| Timing | Sequencing and tempo of movement phases | Hip-shoulder separation in golf swing |
| Range of Motion | Extent of movement through the full intended path | Shoulder external rotation in tennis serve |
| Symmetry | Balance between left and right sides | Weight distribution during bilateral squat |
| Stability | Control and steadiness during movement | Knee valgus control during landing |
Score Interpretation
| Score Range | Interpretation | Recommendation |
|---|---|---|
| 0-39 | Significant form issues detected | Reduce weight/intensity, focus on fundamental movement patterns |
| 40-59 | Below average form with specific corrections needed | Address the top 1-2 corrections before progressing |
| 60-79 | Acceptable form with room for improvement | Safe to train at current intensity while refining technique |
| 80-89 | Good form with minor refinements possible | Focus on consistency and sport-specific optimizations |
| 90-100 | Excellent form across all measured criteria | Maintain current technique, consider increasing difficulty |
Reference Models and Calibration
SportsReflector's scoring is calibrated against reference models derived from multiple sources: expert athlete demonstrations, published biomechanical research, and coaching guidelines from recognized governing bodies. Each sport and exercise has a dedicated reference model that defines the acceptable ranges for each scoring criterion.
Reference models account for natural variation in body proportions. The scoring system evaluates relative joint positions and angles rather than absolute positions, so athletes of different heights, limb lengths, and body types are evaluated on the same scale. For example, squat depth is measured by hip-to-knee angle rather than absolute hip height.
Research supporting the biomechanical criteria used in SportsReflector's models includes work on optimal joint angles in resistance training and studies on biomechanical analysis of sport-specific movements published in peer-reviewed journals.
Confidence Reporting
Every analysis includes a confidence percentage that indicates how reliably the AI detected body joint positions. This is separate from the form score — a high confidence level means the AI is certain about what it detected, while the form score evaluates the quality of what was detected.
| Confidence Level | Meaning | Action |
|---|---|---|
| 85-100% | High confidence — all joints clearly detected | Form score and corrections are reliable |
| 70-84% | Moderate confidence — some joints partially occluded | Form score is approximate; re-record from a clearer angle if possible |
| Below 70% | Low confidence — significant joint detection issues | Re-record with better lighting, angle, or distance |
Validation Process
SportsReflector's scoring models are validated through comparison with expert human evaluations. For each sport and exercise, certified coaches and biomechanists independently score a set of test videos. The AI's scores are then compared against the human expert consensus to measure agreement.
This validation process is ongoing. As the app processes more movements and receives user feedback, the reference models are refined to improve accuracy and reduce edge cases where the AI's assessment diverges from expert opinion. Users can flag analyses they believe are inaccurate, and these flagged cases are reviewed during model updates.
It is important to note that even human experts do not agree perfectly on technique scores. Inter-rater reliability studies in sports coaching show that experienced coaches typically agree within 10-15 points on a 100-point scale. SportsReflector aims to fall within this same range of expert agreement.
What the Score Does Not Measure
Transparency about limitations is essential for responsible AI coaching. SportsReflector's form score evaluates observable biomechanical patterns but does not measure several important factors:
- •Internal forces: The AI cannot measure muscle activation, tendon loading, or internal joint forces. Two movements that look identical externally may involve different muscle recruitment patterns.
- •Pain or discomfort: The app cannot detect whether a movement causes pain. Athletes should not rely solely on form scores to determine exercise safety.
- •Fatigue effects: While the AI can detect form degradation over a set, it cannot measure physiological fatigue or recovery status.
- •Individual anatomy: Some athletes have structural variations (hip socket depth, spinal curvature, limb length ratios) that affect optimal movement patterns. The reference models accommodate common variation but cannot account for all individual differences.
- •Tactical context: In sports like basketball or soccer, the "correct" technique depends on game context (defender position, court spacing). The AI evaluates biomechanics in isolation.
Research References
The following published research informs SportsReflector's methodology and provides context for the accuracy and limitations of AI-powered sports analysis:
Pose Estimation Accuracy
Cao, Z., et al. "OpenPose: Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields." IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. arXiv:1812.08008
MediaPipe Framework
Lugaresi, C., et al. "MediaPipe: A Framework for Building Perception Pipelines." Workshop on Computer Vision for AR/VR, CVPR 2019. arXiv:1906.08172
Biomechanical Analysis in Sport
Colyer, S.L., et al. "A Review of the Evolution of Vision-Based Motion Analysis and the Integration of Advanced Computer Vision Methods Towards Developing a Markerless System." Sports Medicine - Open, 2018. DOI: 10.1186/s40798-018-0139-y
AI in Sports Coaching
Naik, B.T., et al. "A Comprehensive Review of Computer Vision in Sports: Open Issues, Future Trends and Research Directions." Applied Sciences, 2022. DOI: 10.3390/app12094429
Resistance Training Biomechanics
Schoenfeld, B.J. "Squatting Kinematics and Kinetics and Their Application to Exercise Performance." Journal of Strength and Conditioning Research, 2010. PubMed: 20182386
